New hash function alternatives #310

sleeepyjack · 2023-05-19T16:59:28Z

This PR introduces a set of new hash function alternatives, namely

xxhash_32 and xxhash_64: 32- and 64-bit versions of the famous xxHash hash function.
fmix_32 and fmix_64: 32- and 64-bit versions of the Murmur3 integer/avalanche finalizer.

xxhash_32 is now the default hash function for all cuco data structures, due to its better performance in scenarios where the data structure fits into L1$ or L2$. For larger data structures, the performance advantage over Murmurhash3 is negligible.

Benchmark scenario:
Run cuco::static_set::contains() multiple times in a row to ensure the set is resident in the cache. Target occupancy is set to 50%. We vary the table size to see the effects of different cache levels. Ref Time shows the runtime for murmurhash3_32 and Cmp Time shows the time for xxhash_32. All experiments have been conducted on a NVIDIA A100-SXM4-80GB.

Key	Distribution	NumInputs	Ref Time	Ref Noise	Cmp Time	Cmp Noise	Diff	%Diff
I32	UNIQUE	8000	654.880 us	0.44%	658.101 us	0.40%	3.220 us	0.49%
I32	UNIQUE	80000	850.838 us	0.36%	813.983 us	0.40%	-36.854 us	-4.33%
I32	UNIQUE	800000	2.372 ms	2.98%	2.319 ms	0.10%	-52.546 us	-2.22%
I32	UNIQUE	8000000	28.026 ms	0.16%	27.976 ms	0.02%	-50.574 us	-0.18%
I32	UNIQUE	80000000	319.231 ms	0.03%	318.746 ms	0.01%	-484.905 us	-0.15%
I64	UNIQUE	8000	680.446 us	0.40%	646.500 us	0.45%	-33.946 us	-4.99%
I64	UNIQUE	80000	854.186 us	0.28%	827.175 us	0.32%	-27.010 us	-3.16%
I64	UNIQUE	800000	2.628 ms	0.10%	2.523 ms	0.93%	-104.811 us	-3.99%
I64	UNIQUE	8000000	32.144 ms	0.03%	31.949 ms	0.01%	-195.392 us	-0.61%
I64	UNIQUE	80000000	346.017 ms	0.00%	344.892 ms	0.01%	-1124.911 us	-0.33%

In addition, here are some isolated benchmarks calling each hash function in a tight loop on the device:

Hash	Samples	CPU Time	Noise	GPU Time	Noise	Elem/s
`cuco::detail::MurmurHash3_32<int>`	332x	1.520 ms	0.68%	1.510 ms	0.26%	662.091G
`cuco::detail::MurmurHash3_32<long>`	174x	2.891 ms	0.33%	2.882 ms	0.02%	347.000G
`cuco::detail::MurmurHash3_32<key_128_bytes>`	39x	12.910 ms	0.08%	12.900 ms	0.01%	77.521G
`cuco::detail::XXHash_32<int>`	405x	1.245 ms	0.78%	1.236 ms	0.04%	809.233G
`cuco::detail::XXHash_32<long>`	206x	2.443 ms	0.40%	2.434 ms	0.02%	410.887G
`cuco::detail::XXHash_32<key_128_bytes>`	48x	10.444 ms	0.10%	10.434 ms	0.01%	95.839G
`cuco::detail::XXHash_64<int>`	193x	2.614 ms	0.37%	2.604 ms	0.02%	384.051G
`cuco::detail::XXHash_64<long>`	156x	3.231 ms	0.30%	3.221 ms	0.02%	310.452G
`cuco::detail::XXHash_64<key_128_bytes>`	21x	24.833 ms	0.04%	24.822 ms	0.01%	40.286G
`cuco::detail::MurmurHash3_fmix32<int>`	464x	1.089 ms	0.88%	1.080 ms	0.05%	926.083G
`cuco::detail::MurmurHash3_fmix64<long>`	277x	1.819 ms	0.54%	1.809 ms	0.02%	552.668G

Closes #290

benchmarks/defaults.hpp

benchmarks/hash_bench.cu

include/cuco/detail/hash_functions/murmurhash3.cuh

include/cuco/detail/hash_functions/xxhash.cuh

include/cuco/dynamic_map.cuh

include/cuco/hash_functions.cuh

Co-authored-by: Yunsong Wang <[email protected]>

sleeepyjack · 2023-05-22T13:09:29Z

@bdice would you be so kind to leave a review if you have bandwidth?

include/cuco/static_set.cuh

PointKernel

Looks great!

PointKernel

Looks great!

bdice

Awesome work here @sleeepyjack! I see this PR was already merged but I wanted to look it over just so I am familiar with the implementation. I have a few comments here that you can apply in a follow-up PR if you would like. 👍

bdice · 2023-05-22T18:13:58Z

benchmarks/CMakeLists.txt

@@ -78,3 +78,6 @@ ConfigureBench(DYNAMIC_MAP_BENCH
  hash_table/dynamic_map/find_bench.cu
  hash_table/dynamic_map/contains_bench.cu
  hash_table/dynamic_map/erase_bench.cu)
+
+ConfigureBench(HASH_BENCH


Match existing style so this doesn't look like it belongs under the dynamic map benchmarks:

Suggested change

ConfigureBench(HASH_BENCH

###################################################################################################

# - hash benchmarks -------------------------------------------------------------------------------

ConfigureBench(HASH_BENCH

bdice · 2023-05-22T18:15:06Z

benchmarks/hash_bench.cu

+using namespace cuco::benchmark;
+using namespace cuco::utility;


I'd prefer to avoid using declarations. If you write the full types, that will help readers who are unfamiliar with cuco internals (like me) know what namespaces contain the identifiers being used.

Right, we consider benchmarks as user-facing code so we should not make these kinds of shortcuts. Since these namespaces are a bit wordy, how about two-character namespace aliases?

I don't mind the verbosity at all, honestly. 😄 I'd vote cuco::utility over ut or cu or util any day.

bdice · 2023-05-25T22:09:15Z

include/cuco/detail/hash_functions/xxhash.cuh

+    char const* const bytes      = (char const*)&key;      ///< per-byte access
+    uint32_t const* const blocks = (uint32_t const*)&key;  ///< 4-byte word access


Avoid C-style casts.

bdice · 2023-05-25T22:10:20Z

include/cuco/detail/hash_functions/xxhash.cuh

+      }
+    }
+
+    // the following loop is only needed if the size of the key is no multiple of the block size


Suggested change

// the following loop is only needed if the size of the key is no multiple of the block size

// the following loop is only needed if the size of the key is not a multiple of the block size

bdice · 2023-05-25T22:12:51Z

include/cuco/detail/hash_functions/xxhash.cuh

+template <typename Key>
+struct XXHash_32 {
+ private:
+  static constexpr uint32_t prime1 = 0x9E3779B1U;


I have a weak preference for lowercase hex literals / literal suffixes like 0x9e3779b1u because I find them more readable, but that's totally a matter of taste.

bdice · 2023-05-25T22:13:35Z

include/cuco/detail/hash_functions/xxhash.cuh

+    char const* const bytes       = (char const*)&key;      ///< per-byte access
+    uint32_t const* const blocks4 = (uint32_t const*)&key;  ///< 4-byte word access
+    uint64_t const* const blocks8 = (uint64_t const*)&key;  ///< 8-byte word access


Avoid C-style casts.

bdice · 2023-05-25T22:15:07Z

include/cuco/detail/hash_functions/xxhash.cuh

+    char const* const bytes      = (char const*)&key;      ///< per-byte access
+    uint32_t const* const blocks = (uint32_t const*)&key;  ///< 4-byte word access
+
+    uint32_t offset = 0;


Probably should be a std::size_t to match nbytes.

bdice · 2023-05-25T22:15:18Z

include/cuco/detail/hash_functions/xxhash.cuh

+    uint32_t const* const blocks4 = (uint32_t const*)&key;  ///< 4-byte word access
+    uint64_t const* const blocks8 = (uint64_t const*)&key;  ///< 8-byte word access
+
+    uint64_t offset = 0;


Probably should be a std::size_t to match nbytes.

bdice · 2023-05-25T22:23:36Z

tests/utility/hash_test.cu

+    CHECK(h10(k10) == 2031761887105658523ULL);
+  }
+
+  // TODO SECTION("Check if device-generated hash values match the reference implementation.")


Is this still a TODO?

Ha, good catch! Yes, I'll add them in the next PR.

sleeepyjack · 2023-05-25T22:27:58Z

@bdice thank you so much! Reading through your comments, they are all valid and I will address them in a follow-up PR.

bdice · 2023-05-25T22:35:38Z

Also, I didn't leave a comment about this but wanted to share my work: in cudf I chose to use an intrinsic to rotate left:

https://github.com/rapidsai/cudf/blob/37f76c820ddf833a80b4ce706b9c3e84908e51ff/cpp/include/cudf/detail/utilities/hash_functions.cuh#L62-L66

However, it looks like this compiles out to shf.l.wrap.b32 either way so it doesn't matter which you choose.
https://godbolt.org/z/8b199szzK

There is no 64-bit rotate left intrinsic (afaik) so I would not recommend a change here (so that 32-bit and 64-bit rotation functions look the same).

sleeepyjack added 19 commits April 19, 2023 00:28

Add new hasher using XXH64 reference implementation

d0ffc9f

Add benchmark for hash functions

8eff298

Add Murmur3 integer finalizers/hashers

1b2e6ce

Add XXH32 hash function

83ebf3a

Make sure hash computation does not get optimized out benchmark

d43cf3a

Fixes and optimizations for XXH32

66e0887

Run hash func benchmark in a tight loop

16b17b2

Merge remote-tracking branch 'upstream/dev' into xxhash64

f6b102a

Port XXH64 to new design

23e97d5

Add more tests

0833638

Merge remote-tracking branch 'upstream/dev' into new-hashers

01f7120

Add guard around int128 in unit test

8d59538

Move hash function implementations to a separate directory

fb43c0f

Add default hasher

311c831

Roll back default hasher

ff2ce08

Add benchmark for static_set::find table sizes

216064c

Add set contains capacity benchmark

cd7469b

Add xxhash copyright disclaimer

3b58127

Set xxhash as the default hash function

ef5d850

sleeepyjack added helps: rapids Helps or needed by RAPIDS topic: performance Performance related issue Needs Review Awaiting reviews before merging labels May 19, 2023

sleeepyjack requested review from PointKernel and jrhemstad May 19, 2023 16:59

PointKernel reviewed May 19, 2023

View reviewed changes

sleeepyjack and others added 5 commits May 20, 2023 01:26

Apply suggestions from code review

d51a972

Co-authored-by: Yunsong Wang <[email protected]>

[pre-commit.ci] auto code formatting

d72bd38

Undo xxhash as default hasher

d550741

Define nvbench size range for testing small data structures

339f907

Remane fmix functions

8b1cf87

sleeepyjack requested a review from PointKernel May 20, 2023 01:56

Remove unnecessary optimizations for better readability

5d468d5

PointKernel reviewed May 22, 2023

View reviewed changes

include/cuco/static_set.cuh Outdated Show resolved Hide resolved

sleeepyjack added 3 commits May 25, 2023 14:35

Revert xxhash as default hasher in static_set

3d3f730

Hardcode grid size in benchmark

3d05d30

Merge remote-tracking branch 'upstream/dev' into new-hashers

73e8412

PointKernel approved these changes May 25, 2023

View reviewed changes

sleeepyjack merged commit 8e49022 into NVIDIA:dev May 25, 2023

bdice reviewed May 25, 2023

View reviewed changes

sleeepyjack mentioned this pull request Jun 1, 2023

xxhash cleanups #313

Merged

PointKernel mentioned this pull request Jul 5, 2023

[FEA]: Replace cuco::detail::MurmurHash3_32 with a proper public API rapidsai/cugraph#3692

Closed

2 tasks

vyasr mentioned this pull request May 17, 2024

[FEA] Support XXHash64 hashing rapidsai/cudf#12829

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

New hash function alternatives #310

New hash function alternatives #310

sleeepyjack commented May 19, 2023 •

edited

Loading

sleeepyjack commented May 22, 2023

PointKernel left a comment

PointKernel left a comment

bdice left a comment

bdice May 22, 2023

bdice May 22, 2023

sleeepyjack May 25, 2023

bdice May 25, 2023

bdice May 25, 2023

bdice May 25, 2023

bdice May 25, 2023

bdice May 25, 2023

bdice May 25, 2023

bdice May 25, 2023

bdice May 25, 2023

sleeepyjack May 25, 2023

sleeepyjack commented May 25, 2023

bdice commented May 25, 2023 •

edited

Loading

		using namespace cuco::benchmark;
		using namespace cuco::utility;

		char const* const bytes = (char const*)&key; ///< per-byte access
		uint32_t const* const blocks = (uint32_t const*)&key; ///< 4-byte word access

	// the following loop is only needed if the size of the key is no multiple of the block size
	// the following loop is only needed if the size of the key is not a multiple of the block size

New hash function alternatives #310

New hash function alternatives #310

Conversation

sleeepyjack commented May 19, 2023 • edited Loading

sleeepyjack commented May 22, 2023

PointKernel left a comment

Choose a reason for hiding this comment

PointKernel left a comment

Choose a reason for hiding this comment

bdice left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sleeepyjack commented May 25, 2023

bdice commented May 25, 2023 • edited Loading

sleeepyjack commented May 19, 2023 •

edited

Loading

bdice commented May 25, 2023 •

edited

Loading